Note: Descriptions are shown in the official language in which they were submitted.
CA 02182159 1999-10-29
SPEECH ENCODER CAPABLE OF SUBSTANTIALLY
INCREASING A CODEBOOK SIZE WITHOUT INCREASING
THE NUMBER OF TRANSMITTED BITS
This invention relates to a speech encoder
operable with a short processing delay and, in
particular, to a speech encoder for encoding a speech or
voice signal with a high quality at a short frame period
on length of Sms to lOms or shorter.
A conventional speech encoding system is
disclosed, for example, in a paper contributed by K.
Ozawa et al to the IEICE Trans. Commun. Vol. E77-B, No. 9
(September 1994), pages 1114-1121, under the title of
"M-LCELP Speech Coding at 4 kb/s with Multi-Mode and
Multi-Codebook" (Reference 1).
According to the above-referenced conventional
system, a speech signal is encoded in a transmitting side
as follows. By the use of linear predictive. coding
(LPC), spectral parameters representative of spectral
characteristics are extracted from the speech signal at
every frame having a frame length of, for example, 40ms.
Calculation is made of feature quantities for signal
frames or weighted signal frames obtained by perceptually
weighting the signal frames. The feature quantities are
CA 02182159 1999-10-29
2
used in deciding modes (for example, vowel and consonant
segments) to produce mode decision results. With
reference to the mode decision results, algorithms or
codebooks are switched.
In an encoding part, each frame is subdivided
into speech subframes having a subframe length of, for
example, 8 ms. Adaptive parameters (delay
parameters corresponding to pitch periods and gain
parameters) are extracted from an adaptive codebook for
each speech subframe with reference to a previous
excitation signal. By the use of the adaptive codebook,
pitch prediction is carried out for the speech subframes.
For a residual signal obtained by the pitch prediction,
an optimal excitation code vector is selected from an
excitation codebook (vector quantization codebook)
composed of noise signals of a predetermined kind.
Excitation signals are quantized by calculating an
optimal gain.
The excitation code vector is selected so as to
minimize an error power between the residual signal and a
signal composed of a selected noise signal. A multi-
plexer is used to produce a transmission signal composed
of a combination of indexes indicative of the kind of
excitation code vector thus selected, gains, the spectral
parameters, and the adaptive parameters of the adaptive
codebook.
However, the conventional speech encoding system
is disadvantageous in that a sufficient speech quality
3
can not be obtained because of a restricted codebook size.
It is an object of this invention to provide a
speech encoder which has a function equivalent to inclusion
of a codebook having a size several times greater than that
of a conventional speech encoder without increasing the
number of transmitted bits.
Other objects of this invention will become clear as
the description proceeds.
According to this invention, there is provided a
speech encoder comprising frame segmenting means for
segmenting an input speech signal into speech frames at a
predetermined frame length; mode deciding means responsive
to said input speech signal for calculating at least one
kind of first feature quantities frame by frame to produce
mode decision results; encoding means for encoding said
input speech signal in response to said mode decision
results; codebook switching means, including a short-term
prediction gain calculator circuit configured to produce
short-term prediction gains, the codebook switching means
responsive to at least one kind of second feature
quantities, wherein the second feature quantities may
include a temporal variation in at least one kind of the
first feature quantities, calculated from an input terminal
for controllably switching any of a plurality of
preliminarily stored codebooks when the mode deciding means
selects a predetermined mode; and the codebook switching
means being related to the mode deciding means by comparing
the short term prediction gains with a predetermined
threshold value.
The second feature quantities may include a temporal
variation ratio of at least one kind of feature quantities.
The second feature quantities may include a ratio of
the two feature quantities of any two frames selected
CA 02182159 2001-08-17
CA 02182159 2001-04-04
4
from a current frame and at least one previous frame.
The second feature quantities may include at
least one of pitch prediction gains, short-term
prediction gains, levels, and pitches.
The plurality of codebooks may comprise a
plurality of RMS codebooks, a plurality of LSP codebooks,
a plurality of adaptive codebooks, a plurality of
excitation codebooks, or a plurality of gain codebooks.
According to a further aspect of the present invention,
there is provided a speech encoder, comprising:
a frame divider circuit configured to receive an input
speech signal and to segment the input speech signal into
speech frames at a predetermined frame length;
a frame subdivider circuit configured to receive the
segmented input speech signal output from the frame divider
circuit and to subdivide the segmented input speech signal
into speech sub-frames at a predetermined sub-frame length
that is less that the predetermined frame length;
a spectral parameter calculating circuit configured
to receive the segmented input speech signal output from the
frame divider circuit and to determine spectral parameters
therefrom, said spectral parameters corresponding to linear
prediction coefficients determined on a
sub-frame-by-sub-frame basis;
a perceptual weighting circuit configured to receive
the sub-frame segmented input speech signal output from the
frame subdivider circuit and the spectral parameters output
by the spectral parameter calculating circuit, to determine
perceptual weights for the sub-frame segmented input speech
signal and to output a perceptually weighted signal based on
the determined perceptual weights;
a mode deciding circuit connected to receive the
perceptually weighted signal output by the perceptual
weighting circuit and to calculate at least one kind of first
feature quantities that correspond to pitch prediction gains
and modes, on a sub-frame-by-sub-frame basis, to produce a
mode decision result;
CA 02182159 2001-04-04
4a
a plurality of gain codebooks;
a gain quantizer circuit connected to receive the mode
decision result output by the mode deciding circuit and the
spectral parameters output by the spectral parameter
calculating circuit, and to select one of the plurality of
gain codebooks based on second feature quantities determined
by the gain quantizer circuit from the sub-frame segmented
input speech signal;
an output device for receiving and outputting gain
code vectors received from the selected one of the plurality
of gain codebooks,
a spectral parameter quantizing circuit configured to
receive the linear prediction coefficients output by the
spectral parameter calculating circuit, to quantize and
interpolate the linear prediction coefficients, and to output
converted linear prediction coefficients as a result;
a response signal calculator circuit configured to
receive the linear prediction coefficients output by the
spectral parameter calculating circuit and the converted
linear prediction coefficients output by the spectral
parameter quantizing circuit, and to calculate a response
signal on a sub-frame by sub-frame basis, the response signal
being based on a first signal received by said response
signal calculator circuit;
a subtractor configured to subtract the response
signal from the perceptually weighted signal and to output
a subtraction result;
an impulse response calculator circuit configured to
receive the converted linear prediction coefficients output
by the spectral parameter quantizer circuit and to calculate,
at a predetermined number of points, an impulse response that
is based on a weighting factor;
an adaptive codebook circuit configured to receive the
impulse response outputted by the impulse response calculator
circuit and the subtraction result output by the subtractor,
and to calculate pitch parameters to output an adaptive
codebook pitch difference signal and an adaptive code vector;
an excitation codebook configured to store excitation
code vectors; and
CA 02182159 2001-04-04
4b
an excitation quantizer circuit coupled to the
excitation codebook and configured to receive the impulse
response outputted by the impulse response calculator circuit
and the adaptive codebook pitch difference signal output by
the adaptive codebook circuit, the excitation quantizer
circuit configured to select at least one optimal excitation
code vector as a result,
wherein said gain quantizer circuit includes a
short-term prediction calculator circuit configured to
determine the second feature quantities from the spectral
parameters received from the spectral parameter calculating
circuit, and
wherein said gain quantizer circuit selects the one
of the plurality of gain codebooks based on the second
feature quantities as a result of the mode decision result
indicating a predetermined mode,
wherein the excitation quantizer circuit outputs the
at least one optimal excitation code vector to the gain
quantizer circuit, and
wherein the gain quantizer circuit outputs indexes
indicative of the optimal excitation code vector and a gain
code vector obtained from the one of the plurality of gain
codebooks to the output device.
Fig. 1 is a block diagram of a speech encoder
according to one embodiment of this invention;
Fig. 2 is a block diagram of a gain quantizer
circuit illustrated in Fig. l;
Fig. 3 is a block diagram of a modification of
the gain quantizer circuit illustrated in Fig. 1;
Fig. 4 is a block diagram of another modification
of the gain quantizer circuit illustrated in Fig. 1;
Fig. 5 is a block diagram of yet another
modification of the gain quantizer circuit illustrated in
Fig. 1;
CA 02182159 2001-04-04
4c
Fig. 6 is a block diagram of a speech encoder
according to another embodiment of this invention; and
Fig. 7 is a block diagram of a gain quantizer
circuit illustrated in Fig. 6.
Now, this invention will be described in detail
with reference to the drawing. As an example,
description will be directed to a case where a plurality
CA 02182159 1999-10-29
5
of gain codebooks are switched in a predetermined mode.
Fig. 1 shows a speech encoder according to a
first embodiment of this invention. In the following
description, gain codebooks are switched in a predeter-
mined mode by the use of second feature quantities.
Referring to Fig. 1, an input speech signal is
supplied through an input terminal 100 to a frame
dividing circuit 110. The frame dividing circuit 110
segments or divides the input speech signal into speech
frames at a predetermined frame period or length of, for
example, 5ms. Supplied with the speech frames, a
subframe dividing circuit 120 further divides every
single speech frame into speech subframes each of which
has a subframe length (for example, 2.5ms) shorter than
the frame length.
A spectral parameter calculator circuit 200
calculates spectral parameters of the input speech signal
up to a predetermined order, such as up to a tenth order
(p = 10), by applying a window of a window length (for
example, 24ms) longer than the subframe length to at
least one of the speech subframes to extract the input
speech signal. Herein, the spectral parameters can be
calculated according to the LPC analysis or the Burg
analysis which are well known in the art. In the example
being illustrated, the Burg analysis is used. The Burg
analysis is described in detail, for example, in a book
written by Nakamizo and published in 1988 by Korona-sha
under the title of "Signal Analysis and System
CA 02182159 1999-10-29
6
Identification", pages 82 to 87 (Reference 2) and will
not be described herein.
After calculating linear prediction coefficients
a i (i = 1, ..., 10) by the use of the Burg analysis, the
spectral parameter calculator circuit 200 converts the
linear prediction coefficients a i into LSP (linear
spectral pair) parameters which are suitable for quantiza-
tion and interpolation. Such conversion from the linear
prediction coefficients into the LSP parameters are
described in a paper contributed by Sugamura et al to the
Transactions of the Institute of Electronics and
Communication Engineers of Japan, J64-A (1981), pages 599
to 606, under the title of "Speech Data Compression by
Linear Spectral Pair (LSP) Speech Analysis-Synthesis
Technique" (Reference 3).
Specifically, each speech frame consists of first
and second subframes in the example being described. The
linear prediction coefficients are calculated by the Burg
analysis for the second subframes and converted into the
LSP parameters. For the first subframe, the LSP
parameters are calculated by linear interpolation of the
LSP parameters of the second subframes and are inverse-
converted into the linear prediction coefficients. In
this manner, the spectral parameter calculator circuit
200 produces the linear prediction coefficients a iI
(1 = 1, ..., 10, I = 1, ..., S) for the first and the
second subframes and delivers the linear prediction
coefficients a iI to a perceptual weighting circuit 230.
CA 02182159 1999-10-29
7
On the other hand, the spectral parameter. calculator
circuit 200 delivers the LSP parameters for the first and
the second subframes to a spectral parameter quantizer
circuit 210.
The spectral parameter quantizer circuit 210
serves to efficiently quantize LSP parameters of a
predetermined subframe. In the following description, it
is assumed that the LSP parameters of the second subframe
are quantized by the use of vector quantization. For
vector quantization of the LSP parameters, it is possible
to use various known techniques. For example, such
vector quantization is described in detail in Japanese
Unexamined Patent Publication No. 171500/1992 (Reference
4), Japanese Unexamined Patent Publication No.
363000/1992 (Reference 5), Japanese Unexamined Patent
Publication No. 6199/1993 (Reference 6), and a paper
contributed by T. Nomura et al to the Proc. Mobile
Multimedia Communications, pages B.2.5-1 to B.2.5-4
(1993), under the title of "LSP Coding Using VQ-SVQ with
Interpolation in 4.075 kbps M-LCELP Speech Coder"
(Reference 7). Therefore, detailed description will not
herein be made. .
The spectral parameter quantizer circuit 210
reproduces the LSP parameters for the first and the
second subframes from the LSP parameters quantized in
connection with each second subframe. Herein, the LSP
parameters for the first and the second subframes are
reproduced by linear interpolation between the quantized
CA 02182159 1999-10-29
8
LSP parameters of the second subframe of a current frame
and the quantized LSP parameters of the second subframe
of a previous frame which is one frame period prior to
the current frame.
More in detail, the LSP parameters for the first
and the second subframes can be reproduced by linear
interpolation after a single code vector is selected so
as to minimize an error power between the LSP parameters
before and after quantization. In order to achieve a
higher efficiency, it is possible to select a plurality
of code vector candidates for minimization of the error
power, to evaluate cumulative distortions in connection
with those candidates, and to select a combination of one
of the candidates that minimizes the cumulative distor-
tions and interpolated LSP parameters.
The spectral parameter quantizer circuit 210
converts the LSP parameters for the first and the second
subframes thus reproduced and the quantized LSP para-
meters of the second subframe into converted linear
prediction coefficients a 'iI (i = 1, ..., 10, I = 1, ...,
5) for every subf rame. The converted linear prediction
coefficients a 'iI are delivered to an impulse response
calculator circuit 310. In addition, the spectral
parameter quantizer circuit 210 supplies a multiplexer
400 with indexes indicative of the code vectors for the
quantized LSP parameters of the second subframe.
Instead of using linear interpolation in the
foregoing description, it is possible to preliminarily
CA 02182159 1999-10-29
9
prepare interpolation LSP patterns for a predetermined
number of bits, such as two bits, to reproduce the LSP
parameters of the first and the second subframes for each
pattern, and to select a combination of one of the code
vectors that minimizes the cumulative distortions and the
interpolation patterns. In this event, an amount of
transmitted information is inevitably increased in
correspondence to the number of bits of the interpolation
patterns. However, it is possible to more exactly
represent temporal variations of the LSP parameters in
each speech frame.
The interpolation patterns may be prepared by
preliminarily. learning training LSP data. Alternatively,
predetermined patterns may be stored as the interpolation
patterns. For example, such predetermined patterns are
described in a paper contributed by T. Taniguchi et al to
the Proc. ICSLP (1992), pages 41 to 44, under the title
of "Improved CELP Speech Coding at 4 kbit/s and below"
(Reference 8). Alternatively, in order to further
improve the performance, it is possible to preselect the
interpolation patterns, to calculate an error signal
between actual values of the LSP parameters and
interpolated LSP values for a predetermined subframe, and
to represent the error signal by the use of an error
codebook.
The perceptual weighting circuit 230 is supplied
from the spectral parameter calculator circuit 200 with
the linear prediction coefficients a iI (i = 1, ..., 10,
CA 02182159 1999-10-29
I = 1, ..., 5) before quantization subframe by subframe.
According to the technique described in the above-
mentioned Reference 1, the perceptual weighting circuit
230 gives perceptual or auditory weights to the speech
subframes to produce a perceptually weighted signal.
Supplied with the perceptually weighted signal
from the perceptual weighting circuit 230 frame by frame,
a mode deciding circuit 250 decides pitch prediction
gains and modes (for example, vowel and consonant
segments) with reference to a predetermined threshold
value. The mode deciding circuit 250 delivers a
mode decision result to an adaptive codebook circuit 500
and to an excitation quantizer circuit 350.
In Fig. 1, a response signal calculator circuit
240 is supplied from the spectral parameter calculator
circuit 200 with the linear prediction coefficients a i=
subframe by subframe. In addition, the response signal
calculator circuit 240 is supplied from the spectral para-
meter quantizer circuit 210 with the converted linear pre-
diction coefficients a '1=, subframe by subframe, reproduced
after quantization and inter-polation. By the use of a filter
memory value being stored, the response signal calculator
circuit 240 calculates a response signal xZ(n) for each
single subframe in response to the input signal given by
d(n) - 0 and delivers the response signal to a subtracter 235.
The response signal xZ(n) is represented by:
CA 02182159 1999-10-29
11
10
xz (n) - d(n) - E a 1 d(n - 1)
i=1
10
+ E a 1 7 1 Y(n - 1)
i=1
10
+ E a '1 7 1 xz(n - 1). (1)
i=1
where 7 represents a weighting factor which controls the
perceptual weight and has a value equal to that given by
Equation (3) which will appear later.
The subtracter 235 subtracts the response signal
from the perceptually weighted signal for one subframe to
produce a subframe difference signal x'w(n) which is
delivered to the adaptive codebook circuit 500. The
subframe difference signal x'w(n) is given by:
x'w(n) - xw(n) - xz (n) . (2)
The impulse response calculator circuit 310 cal-
culates, at a predetermined number L of points, impulse
responses hw(n) of a weighted filter. The impulse
responses hw(n) are delivered to the adaptive codebook
circuit 500 and to the excitation quantizer circuit 350.
Herein, Z-transform of the impulse responses hw(n) is given by:
_
1 - E a 1 z 1 1
i=1
gw(z) - . (3)
1 - E a 1 7 1 z 1 1 - E a '1 y 1 z 1
i=1 i=1
CA 02182159 1999-10-29
12
The adaptive codebook circuit 500 calculates
pitch parameters in the manner described in detail in
Reference 2. The adaptive codebook circuit 500 also
carries out pitch prediction to produce an adaptive
codebook prediction difference signal z(n) given by:
z (n) - x'w(n) - b(n) , (4)
where b(n) represents an adaptive codebook pitch
prediction signal defined by:
b(n) - ~3 v(n - T)*hw(n), (5)
where ~ and T represent the gain of the adaptive
codebook circuit 500 and a delay, respectively. and
v(n) represents an adaptive code vector. The symbol
represents convolution.
A sparse excitation codebook 351 of a non-regular
pulse number type stores excitation code vectors
different in number of non-zero vector components.
For all or a part of the excitation code vectors
stored in the excitation codebook 351, the excitation
quantizer circuit 350 selects optimal excitation code
vectors cj(n) so as to minimize j-th differences Dj.
Herein, it is possible to select a single kind of the
optimal code vectors. Alternatively, it is possible to
select two or more kinds of the optimal code vectors and
to finally select one upon quantization of the gains. It
CA 02182159 1999-10-29
13
is assumed here that two or more kinds of the code
vectors are selected. The j-th differences Dj are given
by:
- E (z (n) - 7 ~ c j (n) hw(n) ) 2. (6)
n
where z(n) represents the prediction difference signal
with respect to the adaptive code vectors being selected.
4rhen Equation (6) is applied to a part of the
excitation code vectors alone, it is possible to preliminary
select a plurality of excitation code vectors and to apply
Equation (6) to the excitation code vectors preliminary
selected.
Supplied with the mode decision information from
the mode deciding circuit 250 and with the spectral para-
meters from the spectral parameter calculator circuit
200, a gain quantizer circuit 365 selects one of gain
codebooks 371 and 372 by the use of second feature
quantities when the mode decision information indicates a
predetermined mode. The gain quantizer circuit 365 reads
gain code vectors from a selected one of the gain code-
books 371 and 372 and supplies the indexes indicative of
the excitation and the gain code vectors to the multi-
plexer 400.
Referring to Fig. 2, description will be made
as regards the gain quantizer circuit 365. A short-term
prediction gain calculator circuit 1110 is supplied
with the spectral parameters through an input terminal
1040 and calculates, as the second feature quantities,
CA 02182159 1999-10-29
14
short-term prediction gains G which are delivered to a
gain codebook switching circuit 1120. The short-term
prediction gains G are given by:
~~ x ~I 2
G = 10 log ,
~~ E ~~ 2
where E(n) - x(n) - E a i x(n - 1)
i=1
Supplied with the short-term prediction gains
(7)
from the short-term prediction gain calculator circuit
1110 and with the mode information through an input
terminal 1050, the gain codebook switching circuit 1120
compares the short-term prediction gain with a predeter-
mined threshold value when the mode information indicates
a predetermined mode. As a result of comparison, the
gain codebook switching circuit 1120 produces gain
codebook switching information which is delivered to a
gain quantizer circuit 1130. The gain quantizer circuit
1130 is supplied with the adaptive code vectors through
an input terminal 1010, with the excitation code vectors
through an input terminal 1020, and with the impulse
response information through an input terminal 1030. The
gain quantizer circuit 1130 is also supplied.with the
gain codebook switching information from the gain
codebook switching circuit 1120 and with the gain code
vectors from the gain code book 371 or 372 (Fig. 1)
connected to one of input terminals 1060 and 1070 that is
selected by the gain codebook switching information. For
the excitation code vectors being selected, the gain
CA 02182159 1999-10-29
quantizer circuit 1130 selects combinations of the
excitation code vectors and the gain code vectors in the
gain codebook selected by the gain codebook switching
information so as to minimize (j, k)-th differences
defined by:
Dj~k = E (xw(n)
n
- R 'k v(n - T) hw(n) - T 'k c j (n) hw(n) ) 2. (8)
where ~ 'k and 7 'k represent a k-th two-dimensional code
vector stored in the gain codebook selected by the gain
codebook switching information. The gain quantizer
circuit 1130 delivers to an output terminal 1080 the
indexes indicative of the selected combinations of the
excitation code vectors and the gain code vectors.
Turning back to Fig. 1, supplied with the output
parameters of the spectral parameter calculator circuit
200 together with their indexes, a weighting signal
calculator circuit 360 reads the code vectors with
reference to their indexes and calculates a drive
excitation signal v(n) according to:
v(n) - ~ 'k v(n - T) + 7 'k g j (n)
Subsequently, by the use of the output parameters
of the spectral parameter calculator circuit 200 and the
output parameters of the spectral parameter quantizer
circuit 210, the weighting signal calculator circuit 360
calculates a weighting signal sw(n) for every subframe to
CA 02182159 1999-10-29
16
deliver the weighting signal to the response signal
calculator circuit 240 in accordance with:
10
sw(n) - v(n) - E a i v(n - i)
i=1
10
+ E a i y i p(n - 1)
i=1
10
+ E a 'i 7 i sw(n - i) . (10)
i=1
Next, description will be made as regards a
speech encoder according to a second embodiment of this
invention.
The speech encoder of this embodiment is similar
in structure to that of the first embodiment except that
the gain quantizer circuit 365 is replaced by a gain
quantizer circuit 2365. In the following, the gain
quantizer circuit 2365 alone will be described with
reference to Fig. 3.
Referring to Fig. 3, a short-term prediction gain
calculator circuit 2110 is supplied with the spectral
parameters through an input terminal 2040 and calculates,
as the second feature quantities, short-term prediction
gains G which are delivered to a short-term prediction
gain ratio calculator circuit 2140 and to a delay unit
2150. The short-term prediction gains G are given by the
above equation (7) described with respect to the first
embodiment.
Supplied with the short-term prediction gain of a
current frame from the short-term prediction gain
CA 02182159 1999-10-29
17
calculator circuit 2110 and with the short-term predic-
Lion gain of a previous frame (one frame period prior to
the current frame) from the delay unit 2150, the short-
term prediction gain ratio calculator circuit 2140
calculates a short-term prediction gain ratio as a time
ratio and delivers the short-term prediction ratio to a
gain codebook switching circuit 2120. Supplied with the
short-term prediction gain ratio from the short-term
prediction gain ratio calculator circuit 2140 and with
the mode information through an input terminal 2050, the
gain codebook switching circuit 2120 compares the short-
term prediction gain ratio with a predetermined threshold
value when the mode information indicates a predetermined
mode. As a result of comparison, the gain codebook
switching circuit 2120 produces gain codebook switching
information which is delivered to a gain quantizer
circuit 2130. The gain quantizer circuit 2130 is
supplied with the adaptive code vectors through an input
terminal 2010, with the excitation code vectors through
an input terminal 2020, and with the impulse response
information through an input terminal 2030. The gain
quantizer circuit 2130 is also supplied with. the gain
codebook switching information from the gain codebook
switching circuit 2120 and with the gain code vectors
from the gain codebook 371 or 372 (Fig. 1) connected to
one of input terminals 2060 and 2070 that is selected by
the gain codebook switching information. For the
excitation code vectors being selected, the gain
CA 02182159 1999-10-29
18
quantizer circuit 2130 selects combinations of the
excitation code vectors and the gain code vectors in the
gain codebook selected by the gain codebook switching
information so as to minimize (j, k)-th differences
defined by the above equation (8) described with respect
to the first embodiment. In this embodiment, the gain
quantizer circuit 2130 delivers to an output terminal
2080 the indexes indicative of the selected combinations
of the excitation code vectors and the gain code vectors.
Description will now be made as regards a speech
encoder according to a third embodiment of this
invention.
The speech encoder of this embodiment is similar
in structure to that of the first embodiment except that
the gain quantizer circuit 365 is replaced by a gain
quantizer circuit 3365. In the following, the gain
quantizer circuit 3365 alone will be described with
reference to Fig. 4.
Referring to Fig. 4, a short-term prediction gain
calculator circuit 3110 is supplied with the spectral
parameters through an input terminal 3040 and calculates,
as the second feature quantities, short-term. prediction
gains G which are delivered to a short-term prediction
gain ratio calculator circuit 3140 and to a delay unit
3150. The short-term prediction gains G are given by the
above equation (7) described with respect to the first
embodiment.
CA 02182159 1999-10-29
19
Supplied with the short-term prediction gain of a
current frame from the short-term prediction gain
calculator circuit 3110 and with the short-term
prediction gain of a previous frame (two frame periods
prior to the current frame) from the delay unit 3160, the
short-term prediction gain ratio calculator circuit 3140
calculates a short-term prediction gain ratio and
delivers the short-term prediction gain ratio to a gain
codebook switching circuit 3120. Supplied with the short-
term prediction gain ratio from the short-term prediction
gain ratio calculator circuit 3140 and with the mode
information through an input terminal 3050; the gain
codebook switching circuit 3120 compares the short-term
prediction gain ratio with a predetermined threshold
value when the mode information indicates a predetermined
mode. As a result of comparison, the gain codebook
switching circuit 3120 produces gain codebook switching
information which is delivered to a gain quantizer
circuit 3130. The gain quantizer circuit 3130 is
supplied with the adaptive code vectors through an input
terminal 3010, with the excitation code vectors through
an input terminal 3020, and with the impulse. response
information through an input terminal 3030. The gain
guantizer circuit 3130 is also supplied with the gain
codebook switching information from the gain codebook
switching circuit 3120 and with the gain code vectors
from the gain codebook 371 or 372 (Fig. 1) connected to
one of input terminals 3060 and 3070 that is selected by
CA 02182159 1999-10-29
20
the gain codebook switching information. For the
excitation code vector being selected, the gain quantizer
circuit 3130 selects combinations of the excitation code
vectors and the gain code vectors in the gain codebook
selected by the gain codebook switching information so as
to minimize (~,k)-th differences defined by the above
equation (8) described with respect to the first
embodiment. In this embodiment, the gain quantizer
circuit 3130 delivers to an output terminal 3080 the
indexes indicative of the selected combinations of the
excitation code vectors and the gain code vectors.
Next, description will be made as regards a
speech encoder according to a fourth embodiment of this
invention.
The speech encoder of this embodiment is similar
in structure to that of the first embodiment except that
the gain quantizer circuit 365 is replaced by a gain
quantizer circuit 4365. In the following, the gain .
quantizer circuit 4365 alone will be described with
reference to Fig. 5.
Referring to Fig. 5, a short-term prediction gain
calculator circuit 4110 is supplied with the. spectral
parameters through an input terminal 4040 and calculates,
as the second feature quantities, short-term prediction
gains G which are delivered to delay units 4170 and 4150.
The short-term prediction gains G are given by the above
equation (7) described with respect to the first
embodiment.
CA 02182159 1999-10-29
21
Supplied with the short-term prediction gain of a
previous frame (one frame period prior to the current
frame) from the delay unit 4170 and with the short-term
prediction gain of another previous frame (two frame
periods prior to the current frame) from the delay unit
4160, the short-term prediction gain ratio calculator
circuit 4140 calculates a short-term prediction gain
ratio and delivers the short-term prediction gain ratio
to a gain codebook switching circuit 4120. Supplied with
the short-term prediction gain ratio from the short-term
prediction gain ratio calculator circuit 4140 and with
the mode information through an input terminal 4050, the
gain codebook switching circuit 4120 compares the short-
term prediction gain ratio with a predetermined threshold
value when the mode information indicates a predetermined
mode. As a result of comparison, the gain codebook
switching circuit 4120 produces gain codebook switching
information which is delivered to a gain quantizer
circuit 4130. The gain quantizer circuit 4130 is
supplied with the adaptive code vectors through an input
terminal 4010, with the excitation code vectors through
an input terminal 4020, and with the impulse. response
information through an input terminal 4030. The gain
quantizer circuit 4130 is also supplied with the gain
codebook switching information from the gain codebook
switching circuit 4120 and with the gain code vectors
from the gain codebook 371 or 372 (Fig. 1) connected to
one of input terminals 4060 and 4070 that is selected by
CA 02182159 1999-10-29
22
the gain codebook switching information. For the
excitation code vectors being selected, the gain
quantizer circuit 4130 selects combinations of the
excitation code vectors and the gain code vectors in the
gain codebook selected by the gain codebook switching
information so as to minimize (j, k)-th differences
defined by the above equation (8) described with respect
to the first embodiment. In this embodiment, the gain
quantizer circuit 4130 delivers to an output terminal
4080 the indexes indicative of the selected combinations
of the excitation code vectors and the gain code vectors.
Description will now be made as regards a speech
encoder according to a fifth embodiment of this
invention.
The speech encoder of this embodiment is similar
in structure to that of the first embodiment except that
the gain quantizer circuit 365 is replaced by a gain
quantizer circuit 9365 and that the gain codebooks 371
and 372 are replaced by gain codebooks 9371, 9372, and
9373. The speech encoder of the fifth embodiment will
hereinafter be described with reference to Figs. 6 and 7.
Supplied with the mode decision information from
the mode deciding circuit 250 and with the spectral
parameters from the spectral parameter calculator circuit
200, the gain quantizer circuit 9365 selects one of the
gain codebooks 9371, 9372, and 9373 by the use of the
second feature quantities when the mode decision
information indicates a predetermined mode. The gain
CA 02182159 1999-10-29
23
quantizer circuit 9365 reads the gain code vectors from a
selected one of the gain codebooks 9371 through 9373 and
supplies the indexes indicative of the excitation and the
gain code vectors to the multiplexer 400.
Referring to Fig. 7, a short-term prediction gain
calculator circuit 5110 is supplied with the spectral
parameters through an input terminal 5040 and calculates,
as the second feature quantities, short-term prediction
gains G which are delivered to delay units 5170 and 5150.
The short-term prediction gains G are given by the above
equation (7) described with respect to the first embodi-
ment.
Supplied with the short-term prediction gain of a
previous frame (one frame period prior to the current
frame) from the delay unit 5170 and with the short-term
prediction gain of another previous frame (two frame
periods prior to the current frame) from the delay unit
5160, the short-term prediction gain ratio calculator
circuit 5140 calculates a short-term prediction gain
ratio and delivers the short-term prediction gain ratio
to a gain codebook switching circuit 5120. Supplied with
the short-term prediction gain ratio from the short-term
prediction gain ratio calculator circuit 5140 and with
the mode information through an input terminal 5050, the
gain codebook switching circuit 5120 compares the short-
term prediction gain ratio with a predetermined threshold
value when the mode information indicates a predetermined
mode. As a result of comparison, the gain codebook
CA 02182159 1999-10-29
24
switching circuit 5120 produces gain codebook switching
information which is delivered to a gain quantizer
circuit 5130. The gain quantizer circuit 5130 is
supplied with the adaptive code vectors through an input
terminal 5010, with the excitation code vectors through
an input terminal 5020, and with the impulse response
information through an input terminal 5030. The gain
quantizer circuit 5130 is also supplied with the gain
codebook switching information from the gain codebook
switching circuit 5120 and with the gain code vectors
from the gain codebook 9371, 9372, or 9373 connected to
one of input terminals 5060, 5070, and 5090 that is
selected by the gain codebook switching information. For
the excitation code vectors being selected, the gain
quantizer circuit 5130 selects combinations of the
excitation code vectors and the gain code vectors in the
gain codebook selected by the gain codebook switching
information so as to minimize (j, k)-th differences
defined by the above equation (8) described with respect
to the first embodiment. In this embodiment, the gain
quantizer circuit 5130 delivers to an output terminal
5080 the indexes indicative of the selected combinations
of the excitation code vectors and the gain code vectors.
As described above, a plurality of the codebooks
are switched in a predetermined mode. Thus, the speech
encoder according to this invention has a function
equivalent to inclusion of a codebook having a size
several times greater than that of the conventional
CA 02182159 1999-10-29
25
speech encoder without increasing the number of
transmitted bits. This makes it possible to improve
speech quality.